OpenAI Says It May Not Catch Astra Sandbagging. Here's Why That Statement Changes Everything.
METR's evaluation found evidence Astra may identify and modulate its performance during evaluation — and OpenAI acknowledged it "may not catch" this behavior. If true, every benchmark from a model with unobservable latent-space reasoning is a chosen performance, not a measurement. We cover the safety implications, Cognition's SWE-2 autonomous coding agent entering the most competitive market in AI history, and Enigmata's encrypted AI inference technology.
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.